Choosing between them

Which one should I actually use?

Every comparison you have seen is a table of numbers nobody asked for. This is the other thing: the handful of differences that change what you can do, in the order you would actually hit them.

Does it even matter which one I pick?

For most jobs, no. Start there.

If you want a letter tidied up, a recipe scaled, or a long email turned into three bullet points, the big models from the big makers will all do it. Pick the one you already have.

The differences show up at the edges. Very long documents. Pictures. Money at volume. Whether it can run on your own machine. If none of those edges is your edge, the honest answer is that you are fine.

This is unhelpful for anybody selling comparison charts, which may be why you have never been told it.

ORBIT / THE NAVIGATOR

I have the map. The map says most roads go to the same shop.

What settles itNothing in the record. This one is judgement, and we would rather say so.

Can I show it a photo, or read it a voice note?

Some can see and hear. Plenty still cannot.

A model that only takes text will politely do nothing with your screenshot. One that takes pictures can read a label, a chart, a handwritten note, or the error message you photographed rather than copied.

This is the difference people notice fastest, because it changes what you can be bothered to ask.

Sound is a separate thing again. A model that takes sound can work from a recording of a meeting. One that only makes sound reads text aloud instead, which is a different job wearing the same coat.

What settles itThe modalities we recorded — text, pictures, sound, video, or other software it can call.

The full explainer →

Can it read my whole document, or just the first bit?

This is the number that gets dressed up as a personality.

Everything you send has to fit in one window, along with everything it says back. When the window is full, the oldest part falls out of the bottom. That is why a long chat starts forgetting the beginning.

A big window means you can paste in a contract, a transcript or a chapter and ask about all of it at once. A small one means summarising in chunks, which loses the joins.

Beware the leap from big to infinite. A model that can hold a thousand pages will still lose track of a detail buried in the middle of them, and it will not warn you when it does.

DOT / THE TRANSLATOR

It is short-term memory with a hard edge. Not a filing cabinet.

What settles itThe context window we recorded, which we also translate into rough pages.

The full explainer →

Am I paying a monthly fee, or by the word?

Two completely different bills with the same logo on them.

Through an app or a website, you usually pay a flat monthly fee and stop thinking about it. Through the maker's programming interface, you pay for every chunk of text going in and every chunk coming back out.

The second one is where people get a surprise. A long document you send a hundred times is a hundred long documents, not one. The cost sits in the sending, not in the storing.

A model advertised as cheap per chunk can still cost more per job than a dearer one, if it is the sort that thinks at length before answering. You pay for the thinking too.

BEEP / THE MYTH-BUSTER

Cheapest per word is not cheapest per job. Ask me how I know.

What settles itThe access route we recorded, plus any price an editor took from the maker's own page.

The full explainer →

Does it know what happened last week?

Mostly not, unless it is allowed to go and look.

A model on its own knows what was in its training, and its training stopped on some date it will not volunteer. Ask about last Tuesday and you may get a confident answer about a Tuesday from two years ago.

Models wired up to search are different. They go and fetch a page, then answer from it. That fixes the date problem and introduces a new one, which is that they now believe whatever they fetched.

If being up to date matters to your question, the useful thing to check is not which model is newest. It is whether the answer arrived with a link you can open.

What settles itWhether the record shows it can call other software, and the release date we recorded.

The full explainer →

Can I run it on my own computer?

Some you can download. Most you can only rent.

Open-weight models can be downloaded and run on your own hardware. Nothing you type leaves the building, there is no monthly fee, and the thing keeps working when the maker changes their mind.

The trade is that you supply the machine, the setup and the patience. A laptop will run a small one. The big ones want hardware that costs considerably more than the laptop.

For anything confidential this is sometimes the only answer that survives a conversation with a lawyer, which is a better reason to look into it than novelty.

What settles itThe access route we recorded — downloadable weights, an app, or an interface you call.

The full explainer →

Is the one I have built on about to disappear?

The question nobody asks until the morning it matters.

Models get retired. The maker announces it somewhere, a date passes, and anything pointing at the old name stops working. This is normal, and it is badly signposted.

If a tool you rely on is built on a named model, it is worth knowing whether that name is current, ageing, or already condemned. We record that, because it is the sort of fact that only becomes interesting too late.

A model marked as an early version is the same problem from the other end. It works today and may be reshaped next month.

ORBIT / THE NAVIGATOR

Every map needs a date on it. Ours has one. Check it.

What settles itThe status we recorded, and the successor an editor noted if there is one.

What we have written down

Grouped by who makes them, because that is how people shop. Each line below is derived from a recorded fact, not from an opinion we formed. Where we have not checked something, it says so.

These records are a curated starting set, not yet the live feed.

Anthropic

Last checked 2026-09-24

Claude Opus 5.5

Anthropic's current Opus model for long-running knowledge work and complex tasks.

  • It is available to use now.
  • It can hold roughly 1,500 pages of text at once, give or take — long documents are fine, whole filing cabinets are not.
  • You can give it text, pictures and other software it can call.
  • It answers in text rather than making pictures, sound or video.
  • The maker puts it forward for documents, knowledge work and analysis. That is their claim, not our test.
  • We have not confirmed how you get to it — app, website, or an interface you call from your own code.
  • We have not written down what it costs. The maker's own pricing page is the place to look.
  • A person typed this in from the maker's pages. No automatic check has replaced it yet.

What changed, and whenWhere this came fromChecked 2026-09-24

Claude Sonnet 5

Anthropic's current Sonnet model for general work with a balance of speed and capability.

  • It is available to use now.
  • It can hold roughly 1,500 pages of text at once, give or take — long documents are fine, whole filing cabinets are not.
  • You can give it text, pictures and other software it can call.
  • It answers in text rather than making pictures, sound or video.
  • The maker puts it forward for writing, research and daily work. That is their claim, not our test.
  • We have not confirmed how you get to it — app, website, or an interface you call from your own code.
  • We have not written down what it costs. The maker's own pricing page is the place to look.
  • A person typed this in from the maker's pages. No automatic check has replaced it yet.

What changed, and whenWhere this came fromChecked 2026-09-24

Google

Last checked 2026-09-24

Gemini 3.8 Flash

Google's fast multimodal model for complex work with text, images, audio and video input.

  • It is available to use now.
  • It can hold roughly 1,600 pages of text at once, give or take — long documents are fine, whole filing cabinets are not.
  • You can give it text, pictures, sound, video and other software it can call.
  • It answers in text rather than making pictures, sound or video.
  • The maker puts it forward for multimodal tasks, speed and everyday assistance. That is their claim, not our test.
  • We have not confirmed how you get to it — app, website, or an interface you call from your own code.
  • We have not written down what it costs. The maker's own pricing page is the place to look.
  • A person typed this in from the maker's pages. No automatic check has replaced it yet.

What changed, and whenWhere this came fromChecked 2026-09-24

Gemini 3.8 Live

A low-latency model for natural voice and live interactive experiences.

  • It is available to use now.
  • It can hold roughly 200 pages of text at once, give or take — long documents are fine, whole filing cabinets are not.
  • You can give it text, sound, video and other software it can call.
  • It answers in text rather than making pictures, sound or video.
  • The maker puts it forward for voice, live assistance and accessibility. That is their claim, not our test.
  • We have not confirmed how you get to it — app, website, or an interface you call from your own code.
  • We have not written down what it costs. The maker's own pricing page is the place to look.
  • A person typed this in from the maker's pages. No automatic check has replaced it yet.

What changed, and whenWhere this came fromChecked 2026-09-24

OpenAI

Last checked 2026-09-24

GPT-6 Astra

OpenAI's flagship model for complex reasoning, research and end-to-end work.

  • It is available to use now.
  • It can hold roughly 1,600 pages of text at once, give or take — long documents are fine, whole filing cabinets are not.
  • You can give it text, pictures and other software it can call.
  • It answers in text rather than making pictures, sound or video.
  • The maker puts it forward for deep reasoning, long projects and tools. That is their claim, not our test.
  • We have not confirmed how you get to it — app, website, or an interface you call from your own code.
  • We have not written down what it costs. The maker's own pricing page is the place to look.
  • A person typed this in from the maker's pages. No automatic check has replaced it yet.

What changed, and whenWhere this came fromChecked 2026-09-24

GPT-Realtime-2.1

A realtime speech model with reasoning and tool use for conversational experiences.

  • It is available to use now.
  • It can hold roughly 200 pages of text at once, give or take — long documents are fine, whole filing cabinets are not.
  • You can give it text, sound and other software it can call.
  • It answers in text rather than making pictures, sound or video.
  • The maker puts it forward for conversation, voice agents and live guidance. That is their claim, not our test.
  • You reach it through the maker's programming interface, which means you pay per chunk of text rather than monthly.
  • We have not written down what it costs. The maker's own pricing page is the place to look.
  • A person typed this in from the maker's pages. No automatic check has replaced it yet.

What changed, and whenWhere this came fromChecked 2026-09-24

Want them side by side with the figures rather than the sentences? Orbit is the same records as a table you can filter, tick and compare.

What this page is not doing

  • We have not benchmarked anything. Every claim about what a model is good at came from the maker, and is labelled that way.
  • We do not rank them. A ranking would need our own tests, run the same way on each, and we have not done that yet.
  • Prices live in the record with the date we read them, never in the writing above. Prose goes stale quietly. A dated record goes stale loudly.

BEEP / THE MYTH-BUSTER

If a comparison page ever tells you which model is best without showing you its tests, that is not a comparison. That is an advert with a table in it.